core(OpenCL): optimize convertTo() with CV_16F (convertFp16() replacement) #24918
relates #24909
relates #24917
relates #24892
Performance changes:
- [x] 12700K (1 thread) + Intel iGPU
|Name of Test|noOCL|convertFp16|convertTo BASE|convertTo PATCH|
|---|:-:|:-:|:-:|:-:|
|ConvertFP16FP32MatMat::OCL_Core|3.130|3.152|3.127|3.136|
|ConvertFP16FP32MatUMat::OCL_Core|3.030|3.996|3.007|2.671|
|ConvertFP16FP32UMatMat::OCL_Core|3.010|3.101|3.056|2.854|
|ConvertFP16FP32UMatUMat::OCL_Core|3.016|3.298|2.072|2.061|
|ConvertFP32FP16MatMat::OCL_Core|2.697|2.652|2.723|2.721|
|ConvertFP32FP16MatUMat::OCL_Core|2.752|4.268|2.662|2.947|
|ConvertFP32FP16UMatMat::OCL_Core|2.706|2.601|2.603|2.528|
|ConvertFP32FP16UMatUMat::OCL_Core|2.704|3.215|1.999|1.988|
Patched version is not worse than convertFp16 and convertTo baseline (except MatUMat 32->16, baseline uses CPU code+dst buffer map).
There are still gaps against noOpenCL(CPU only) mode due to T-API implementation issues (unnecessary synchronization).
- [x] 12700K + AMD dGPU
|Name of Test|noOCL|convertFp16 dGPU|convertTo BASE dGPU|convertTo PATCH dGPU|
|---|:-:|:-:|:-:|:-:|
|ConvertFP16FP32MatMat::OCL_Core|3.130|3.133|3.172|3.087|
|ConvertFP16FP32MatUMat::OCL_Core|3.030|1.713|9.559|1.729|
|ConvertFP16FP32UMatMat::OCL_Core|3.010|6.515|6.309|4.452|
|ConvertFP16FP32UMatUMat::OCL_Core|3.016|0.242|23.597|0.170|
|ConvertFP32FP16MatMat::OCL_Core|2.697|2.641|2.713|2.689|
|ConvertFP32FP16MatUMat::OCL_Core|2.752|4.076|6.483|4.191|
|ConvertFP32FP16UMatMat::OCL_Core|2.706|9.042|16.481|1.834|
|ConvertFP32FP16UMatUMat::OCL_Core|2.704|0.229|15.730|0.176|
convertTo-baseline can't compile OpenCL kernel for FP16 properly - FIXED.
dGPU has much more power, so results are x16-17 better than single cpu core.
Patched version is not worse than convertFp16 and convertTo baseline.
There are still gaps against noOpenCL(CPU only) mode due to T-API implementation issues (unnecessary synchronization) and required memory transfers.
Co-authored-by: Alexander Alekhin <alexander.a.alekhin@gmail.com>
* Replaced most remaining sprintf with snprintf
* Deprecated encodeFormat and introduced new method that takes the buffer length
* Also increased buffer size at call sites to be a little bigger, in case int is 64 bit
* Support cl_image conversion for CL_HALF_FLOAT (float16)
* Support cl_image conversion for additional channel orders:
CL_A, CL_INTENSITY, CL_LUMINANCE, CL_RG, CL_RA
* Comment on why cl_image conversion is unsupported for CL_RGB
* Predict optimal vector width for float16
* ocl::kernelToStr: support float16
* ocl::Device::halfFPConfig: drop artificial requirement for OpenCL
version >= 1.2. Even OpenCL 1.0 supports the underlying config
property, CL_DEVICE_HALF_FP_CONFIG.
* dumpOpenCLInformation: provide info on OpenCL half-float support
and preferred half-float vector width
* randu: support default range [-1.0, 1.0] for float16
* TestBase::warmup: support float16
Fix dynamic loading of clBLAS and clFFT (formerly, clAmdBlas and clAmdFft)
* Fix dynamic loading of clBLAS and clFFT
* Update filenames and function names for clBLAS (formerly, clAmdBlas)
* Update filenames and function names for clFFT (formerly, clAmdFft)
* Uncomment teardown of clFFT; tear down clFFT in same way as clBLAS
* Fix generators for clBLAS and clFFT headers
* Update generators to parse recent clBLAS and clFFT library headers
* Update generators to be compatible with Python 3
* Re-generate OpenCV's clBLAS and clFFT headers
* Update function calls to match names in newly generated headers
* Disable (and comment on) teardown code for clBLAS and clFFT
* Renaming *clamd* files
* Renaming *clamdblas* files to *clblas*
* Renaming *clamdfft* files to *clfft*
* Update generator for CL headers
* Update generator to be compatible with Python 3
- follows iso c++ guideline C.44
- enables default compiler-created constructors to
also be noexcept
original commit: 77e26a7db3
- handled KernelArg, Image2D
